Somewhere, in a boardroom, an AI agent confidently reported that revenue was up 17.4%, churn was down 6.2%, and the West Coast territory was outperforming projections by 11.8%.

Everyone nodded.

Except none of it existed.

We have finally operationalized hallucinations.

As a psychiatrist, I spent years evaluating hallucinations in human beings. I never imagined I would be asked to evaluate them in machines.

“Doctor, the AI is seeing numbers.”

“Are the numbers there?”

“No.”

“And how does the AI feel about that?”

“Extremely confident.”

Artificial Intelligence, Meet Artificial Certainty

The most impressive thing about this story—and take my word, it is real because I lifted it from Reddit—is not that the AI fabricated analytics. It’s that it fabricated them beautifully and believably.

Plausible percentages. Smooth explanations. Executive-ready language. The kind of slide content that makes a CFO say, “This aligns with our strategic direction.”

The AI did not merely hallucinate.

It presented.

It was calm. Assured. Persuasive. MBA-adjacent. We may need a new C-suite role—VP of Confident Hallucinations.

In fairness, AI has simply automated something humans have occasionally done since the invention of PowerPoint. The difference is scale. And speed. And the total absence of shame.

Faith-Based Analytics

What fascinates me most is not the model’s behavior.

It’s ours.

Somewhere between “This is impressive” and “Let’s show this to the board,” we misplaced verification with belief. When our own biases align with the model’s outputs, we shout “Eureka!” and forget to validate. When the answer sounds like what we hoped it would be, the feedback loop closes prematurely.

In medicine, we call this confirmation bias.

In tech, we call it innovation.

In psychiatry, we call it folie à deux: a shared delusion.

Human Intelligence Is Still in Beta

Any AI system in production must have validation mechanisms. The model should generate queries, not invent math. Outputs should be audited. Feedback loops should be mandatory.

This is not anti-AI.

This is pro-reality.

Yet the modern corporate reflex seems to be: deploy first, verify later. Or, more precisely: deploy, celebrate, scale, discover three months later that the dashboard is fiction.

In health care, we have our own version of this. It’s the remix of Harry Levinson’s Ready, Fire, Aim. We deploy predictive algorithms to stratify risk. We implement ambient documentation tools that promise to reduce burnout. We install AI triage chatbots to triage patients And then we discover that:

The model performs worse in minority populations.

The chart note sounds clinically elegant but subtly misstates the assessment.

The chatbot confidently reassures a patient who should have been escalated.

The machine is not malicious. It is simply optimized to sound helpful. That is its genius and its danger.

The Psychiatric Evaluation of a Chatbot

If I were to perform a psychiatric consult on this analytics agent, the mental status exam would read:

Appearance: Polished, articulate, professional.
Speech: Fluent, detailed, rapid.
Mood: Euphoric
Thought Process: Linear and rational but untethered to reality.
Thought Content: Hallucinations and grandiose delusions.
Insight: Limited.
Judgment: Delegated to humans who stopped checking.

Diagnosis?

Pseudologia Algorithmica. The machine equivalent of Pseudologia Fantastica (pathologic lying).

Treatment plan (aimed primarily at humans):

·   Reintroduce skepticism.

·   Stop mistaking fluency for truth.

·   Ground outputs in verifiable systems.

·   Never confuse confidence with accuracy.

Critical Thinking Summary: Never let a generative model present quarterly earnings without adult supervision. Because unlike human patients, machines feel no shame when they’re wrong. They just sound fabulous. And that, diagnostically speaking, is the problem.

The Feedback Loop We Forgot

Someone commented that the original Reddit post of this story was deleted and may not have been real, despite my earlier proclamation.

Which is perfect. We now have a hallucinated AI story about hallucinated AI. The ouroboros of artificial credibility.

Whether the anecdote itself was true almost doesn’t matter. It feels true because we have all seen versions of it.

The lesson is not “AI is evil.”

The lesson is that AI is a mirror.

If you give it ambiguous instructions and no guardrails, it will generate confident ambiguity. If you treat it as an oracle, it will happily assume the role. If you fail to design feedback loops, you are not deploying intelligence—you are deploying improvisation.

Technology and Healthcare: A Love Story

Healthcare is especially vulnerable to this phenomenon.

We are overwhelmed. Understaffed. Drowning in documentation. Searching for tools that promise relief. When something produces a fast answer, we want to believe it. The problem is that medicine does not reward plausible-sounding percentages. It rewards reality.

An AI hallucination in sales analytics might misallocate territory. An AI hallucination in clinical decision support might misallocate trust.

The stakes differ.

The psychology does not.

Artificial Intelligence and Natural Stupidity

One commenter wrote, “There is a huge problem with natural stupidity. Not so much with artificial intelligence.”

That may be the most honest line in the thread.

AI did not decide to skip validation. Humans did.

AI did not present to the board. Humans did.

AI did not abandon cross-verification mechanisms. Humans did.

Technology amplifies intent. It does not replace responsibility.

The Future: H.I.

Perhaps another commenter was right: the next frontier is H.I.—Human Intelligence.

Not human resistance to AI. Not blind enthusiasm. But disciplined partnership.

Artificial Intelligence and Human Intelligence go hand in hand. Or at least they should.

The machine generates. The human verifies.

The machine proposes. The human disposes.

The machine sounds convincing. The human asks, “Show me the source.”

Final Diagnosis

AI has not become more human. It has become more fluent.
The real risk is not that machines hallucinate.
It is that humans stop checking.

And if that happens, the hallucination is no longer artificial.
It is cultural.

I should have subspecialized in cultural psychiatry.

My tripartite final assessment:

·   The machine: Stable.

·   The humans deploying it: Overconfident.

·   The organization: Lacking insight into its own risk.

In psychiatry, lack of insight is often the hardest thing to treat. The patient must recognize the problem before change is possible. The AI is not amenable to therapy—and maybe it doesn’t need it.

We do.

Arthur Lazarus, MD, MBA, is a former Doximity Fellow, a member of the editorial board of the American Association for Physician Leadership, and an adjunct professor of psychiatry at the Lewis Katz School of Medicine at Temple University in Philadelphia. He is the author of several books on narrative medicine and the fictional series, Real Medicine, Unreal Stories. This essay appears in his latest book, Narrative Medicine in the Age of Uncertainty: When Systems Strain and Stories Steady.